Improving hydrological model identifiability by driving calibration with stochastic inputs

For a long time, the classical problem of identifying the optimal modeling structure and/or parameters has been based on the calibration-validation norm, originating from the iconic split-sample scheme by Vit Klemeš and subsequently evolved in several ways. A common feature of such approaches is their dependence on the length and representativeness of the available historical data. This introduces several questions since the derived parameters are selected on the basis of a specific subset (or more generally subsets) of data, while the rest of data is used to evaluate the predictive capacity of the calibrated model. To address this shortcoming, we propose a novel and conceptually simple approach driven by the well-known stochastic simulation paradigm. The method builds upon the idea of calibrating hydrological models using alternative, yet probabilistically consistent, stochastically generated data. Decoupling this way, the available historical data now become the basis to generate synthetic input data, as well as for model validation and parameter uncertainty assessment. One main advantage is embedding the stochasticity of real-world drivers (rainfall, evapotranspiration) and responses (runoff) and thus their hydrological uncertainty. Another advantage is identification of stable and robust models since the calibration procedure is performed using long enough time series that reproduce important stochastic and probabilistic properties that are associated with the changing climate (e.g., long-term persistence) that are generally hidden in the short historical samples. Identifying this way, the derived parameters are optimal not only for the historical data set, but for any alternative plausible realization of the modeled processes.